With the benefit of four months’ or so reflection, it’s a bit easier to make a more rational assessment of the performances of the MoS menagerie of models in 2018.
Here’s where I’ve landed on a few of the key models (see this post from last year for detailed performance data).
The MoS twins - MoSSBODS and MoSHBODS - did well enough as head-to-head, margin and probability forecasters, though MoSHBODS’ margin forecasting performance was disappointing. It finished with a Mean Absolute Error (MAE) about 1 point per game worse than the TAB bookmaker’s, and ahead of only C_Marg.
In 2017, however, it topped the Margin Predictor table and finished about 0.4 points per game ahead of the TAB Bookmaker, so I’ll need another season or two of 2018-like under-performance before I pop the hood on the underlying MoSHBODS algorithm and consider significant changes. Real-world data analytics and predictive modelling is often about responding proportionately to unexpected and unfavourable variability.
MoSSBODS did an admirable job informing Investors’ Line betting, and will back up to do that again this year. It also performed well as a margin predictor and as a head-to-head probability predictor.
MoSHBODS did less well informing Investors’ Head-to-Head betting, though a case can be made that the Fund’s Kelly-staking strategy contributed at least partly to that outcome. More on this later.
The ChiPS-based forecasters - C_Marg and C_Prob - had horrible years (again), which once again had me seriously considering removing ChiPS from the forecaster pool, but I eventually decided to let it run, after re-optimising it, for one more season. More too on this later.
The remaining bookie-based and other models, which I include solely as benchmark models for those just mentioned, I think provided enough interest last season to justify their continuation. ENS_Linear, for example, topped the Head-to-Head Tipping chart and finished only 0.4 points per game behind the TAB Bookmaker in margin prediction.
The MARS Team Rating System, which powers ENS_Linear, continues to provide useful Elo-style team strength estimates, which can be directly compared to those of the more complex MoSBODS and MoSHBODS Systems.
EXISTING FORECASt MODELS
The ENS_Linear model, which I just alluded to, has again been recalibrated, this year using only the data for seasons 2009 to 2018 and the same three base learners. More details on this model can be found in this post from last season.
I’ve also spent some time in the off-season reviewing the calibration of the MoSHBODS Team Rating System. The chart below shows the win percentage for home teams with a given MoSHBODS expected margin (the black dots), a loess smoothed version of that data (the blue line), and the average estimated victory probability of home teams with a given MoSHBODS expected margin (the red dots).
This chart suggests that MoSHBODS slightly underestimates home teams’ victory chances in games where their expected margin is in the range from about a 50-point loss to about a 10-point win. Also hinting at a possible relative underperformance in this range is the fact that MoSHBODS’ MAE for these games across the period from 2000 to 2018 is about 0.5 points higher than for games outside that range (though this difference disappears if we exclude games from 2000 to 2006). I could not, however, find any simple adjustment that materially improved MoSHBODS’ overall log probability score, so I’ve left MoSHBODS and the process for converting its expected margins to probabilities unchanged for season 2019 (see this post from last year for details of how this is done).
MoSHBODS’ margin prediction will continue to be the difference between its home and away team score predictions plus the 2-point bias adjustment, and its head-to-head tip will be the home team if this adjusted margin is positive, and the away team otherwise.
The MoSSBODS Team Rating System will be used analogously to MoSHBODS to also provide Head-to-Head, Margin and Probability forecasts.
(It’s interesting to note that MoSSBODS underestimates to a slightly greater extent the victory chances of what it considers to be home team underdogs or narrow favourites, as we can see from the chart below. It also has an MAE 0.6 points higher for games where its expected margin lies between -50 and +10 points, a difference that also disappears if we exclude seasons 2000 through 2006).
Note that neither MoSHBODS nor MoSSBODS have been reoptimised for 2019, except to the extent that the Venue Performance Values will have altered on the basis of all of the games played in 2018.
So, what about ChiPS then? After topping all three forecasting tables in 2015, it’s had three lacklustre seasons, the most abject of which came last season where it finished 9 tips behind the TAB bookmaker on head-to-head results, 2 points per game behind on MAE for margin prediction, and a long way behind on log probability score for probability estimation.
Chances are I’ll regret not ditching it this season but, in a final attempt to honour the memory of the handsome if dopey late dog it was named after, I’ve reoptimised ChiPS one more time, this year based on the results from 2009 to 2018, and including about a half dozen more combinations of home team and venue for which HGAs have been estimated. The criterion for including a pairing of team and venue in those HGA estimates was that at least 10 games involving that pairing were played across the 10-year period.
So, in total, there will again be nine traditional Head-to-Head Tipsters, 10 Margin Predictors, and six Head-to-Head Probability Predictors, as listed in the table below.
(You can find out more about some of the forecasters in this blog from 2016 and, if you're new, more about how it all works here on MoS in this post from 2015, although many of the statistical models discussed there are no longer in use.)
Ah, but what about the MoS Funds?
MoS will again be wagering in the Head-to-Head, Line and Over/Under markets this season, albeit it with different rules of engagement for the Head-to-Head and Over/Under Funds, and with different portfolio weights for the three Funds.
HEAD-TO-HEAD FUND (WEIGHTING 25%)
The Head-to-Head Fund lost about 17% of its value across the 2018 season after recording a -7% ROI on a 2.5 turn, although it was fractionally in profit at the end of Round 15 before shedding a startling 21% over the next 7 rounds. That season-long performance was starkly different to the Fund in 2017 where it finished up by about 62% with a 21% ROI on a 2.9 turn.
This year we’ll be down-weighting the Fund and making the following changes to the way it works:
Wagers will only be made on home teams and when the estimated edge is over 5% (ie Price x Estimated Probability > 105%).
A fixed-return staking strategy will be used such that, should a wager be successful, the Fund will increase in value by 2% of its original size.
Achieving the second outcome will require that wagers be sized as follows:
Wager Size (as % of original funds) = 1/(50*(Price-1))
Home-team-only was the norm for MoS head-to-head wagering in most seasons prior to 2017, and the return to it this year is partly a result of the apparent underestimation by MoSHBODS of the victory chances of home team underdogs revealed in that earlier chart.
The move away from Kelly-staking is an acknowledgement that, absent the services of a very well-calibrated algorithm, accurately estimating your edge is impossible, and assuming that you can estimate it accurately is especially fraught for short-priced favourites.
For example, if you’ve estimated that a $1.30 favourite has an 85% chance of victory, then a full Kelly stake would have you betting (85% x 1.3 - 1)/(1.3 - 1), or 35% of your bankroll. If that team’s actual chances were 80%, then the correct Kelly stake would be (80% x 1.3 - 1)/(1.3 - 1) or just 13.3% of your bankroll, which is less than 40% of the wager you’d actually make.
Fixed-return staking as I’m applying it here is, broadly, agnostic about the magnitude of the estimated edge, the exception being that we’re only wagering at all when we estimate the edge exceeds the threshold value of 5%. Once we’re over that threshold, however, the bet size is fixed so that, for example, the wager where the estimated edge is 6% is no different to that where it’s estimated to be 20%.
This strategy would have yielded a 7.7% return in 2018 (6.5% ROI on a 1.2 turn), and a 15.9% return in 2017 (8% ROI on a 2.0 turn. In 2016, when MoSSBODS determined the Head-to-Head Fund’s wagers, it would have resulted in a very small loss of 0.3%, which is still much better than the 9.2% loss actually recorded by MoSSBODS.
As always, previous performance does not guarantee future returns.
LINE FUND (WEIGHTING 50%)
Line wagering has been consistently profitable here on MoS, returning profits in 2008, 2010, 2011, 2013, 2014, 2016, 2017 and 2018 (and only small losses or break even results in 2007, 2009 and 2012). A 20% loss in 2015 is the only significantly bad result across the 12 seasons from 2007 to 2018.
Last year, line wagering was informed by MoSSBODS, and saw profits in all but nine rounds. The season finished with the Line Fund producing about an 18% profit from a 13% ROI on a 1.4 turn. Of the 85 bets placed, 53 (62%) were winners.
This year, the Line Fund will rely on MoSSBODS again, and will still be permitted to wager on home and on away teams. Also as per last year, it will place a line bet on a team proportional to the size of its measured 'advantage', only when the handicap set by the TAB bookmaker is assessed by MoSSBODS as being in error by more than 8 points. History suggests that MoS Funds are far better at estimating their edge in the line market than in the head-to-head market, and prices are typically around $2, so the issue with short-priced favourites and miscalibration does not apply.
Bet size, as a percentage of initial Funds, will again be determined as:
Bet Size = (Assessed Margin + Handicap Offered - 8)/300
provided Assessed Margin + Handicap Offered > 8
So, for example, if the Home team is giving 12.5 start and MoSSBODS assesses them as winning by 33 points, the bet would be (33 - 8 - 12.5)/300 = 4.2% of initial Funds.
Instead, if the Away team is receiving 34.5 start and MoSSBODS assesses them as losing by only 12 points, the bet would be (34.5 - 8 - 12)/300 = 4.8% of initial Funds.
The Line Fund will be up-weighted this year to represent 50% of the Overall Portfolio.
Using the last few seasons as a rough guide, I would expect the Fund to wager on about 50-60% of the games (a little over half the time on home teams) with an average bet size of about 1.75% to 2% of the Fund, on which basis the expected turn would be about 1.8 to 2.5.
OVER/UNDER FUND (25%)
The Over/Under Fund disappointed last year and made profits in only 8 of the 25 rounds in which it chose to bet. Overall, it collected only 50 of 105 wagers and finished down over 22% on the season from a -11% ROI on a 2.1 turn.
It was clear that, on a number of occasions, the ignoring of forecast inclement weather led to ill-advised overs wagers early in the week against Totals that almost certainly included a discount for dampness. In light of this, for season 2019, no overs bets will be placed on games where rain is forecast for game day.
The Fund will continue to use MoSSBODS for its wagering advice and will again require a 6 point minimum overlay before wagering.
We should expect this Fund to wager in about 60% to 70% of games, or around 125 to 150 games across the season. Roughly 50% to 60% of those wagers are likely to be unders bets. Each wager will be 2% of initial Funds, so we can expect a turn of around 2 to 3.
All three Funds will wager in whichever of the TAB or Easybet markets are offering the most attractive prices and are accepting my bets and will be placed:
For head-to-head and line bets: as soon as possible after the Line markets are posted on both TAB and Easybet
For over/under bets: as soon as possible after the Totals markets are posted on both TAB and Easybet
The biggest change for 2019 will be the introduction of player-based forecasts, drawing on the analysis I did in this post over at the Statistical Analysis blog.
Simply put, we’ll be estimating the value of teams’ named squads in terms of net scoring ability and using this information, along with MoSHBODS forecasts to produce a hybrid forecast as follows:
Predicted Home Team Margin = 3.06 + 0.72 x MoSHBODS Expected Margin + 2.42 x Difference in Estimated Player Values
Player values are proxied by their forecasted SuperCoach score for their next game.
For now, I’m going to refer to this as the MoSHPlay forecast.
(At some point I might also build a Totals model incorporating player data).
This change will obviously require me to prepare forecasts - and update blogs - later in the week than has been the norm here on MoS. As a result, you’ll need to be checking back much closer to game time to get my latest and final forecasts.
On which topic …
The following is the rough schedule I plan to follow this season (client work permitting):
Monday or Tuesday night: blog with week’s head-to-head, margin and head-to-head probability forecasts, plus Head-to-Head and Line wagers assuming availability of markets.
Tuesday or Wednesday night: blog with latest team rating data
Wednesday or Thursday night: blog with week’s over/under wagers
Thursday night (and otherwise, as required due to team announcements): blog with week’s player-based forecasts
Sunday night (except in rounds with Monday night games): blog with results for the round, plus team dashboard data
At some point in the season I’ll also start projecting the remainder of the season, though this is unlikely to start before the end of, at least, Round 4. Prior to that I think the uncertainty in the estimates of team abilities and the resulting (potentially unmodelled) variability in the results so great as to render the projections broadly useless..
And, finally, I plan also to blog occasionally about any interesting statistical analyses that I get the time to conduct.
It’s going to be a very busy year. Hope you can stay for the ride.